Hi, I’m
Naman Tiwari.

I want to become an AI-focused backend engineer. Right now I’m learning backend development, and I’m looking for an internship where I can build real things with a team.

  • 3rd-year B.Tech student
  • Data Science and AI
  • Open to internships
Contact me → View resume
Portrait of Naman Tiwari

Projects

Clinical Co-Pilot answering a question about an ECG report, with a verified badge and a source list

Clinical Co-Pilot

Multimodal RAG assistant for patient records

Answers a clinician’s question about one patient and cites its sources. A five-step LangGraph pipeline mixes keyword and vector search, reranks the results, keeps each patient’s records separate, and checks every number in an answer against the sources. RAGAS on 40 generated questions: context precision 0.98, recall 0.87, faithfulness 0.81.

  • Python
  • LangChain
  • LangGraph
  • Pinecone
  • Cohere
  • Gradio
See more ↗ Live demo ↗
Voice Shield live dashboard with an impersonation risk gauge and a banking call transcript

Voice Shield

Real-time voice-clone detection for banking calls

A FastAPI and WebSocket service that scores live audio with a fine-tuned AASIST model and decides whether to approve, ask for an OTP or freeze a simulated account. Fine-tuning cut the equal error rate from 46.0% to 13.5% on held-out speakers. A research prototype, not a production fraud system.

  • Python
  • PyTorch
  • FastAPI
  • WebSockets
  • Docker
See more ↗ Live demo ↗
Fraud detection dashboard with a threshold slider and an uploaded CSV of transactions

Fraud Detection

Credit card fraud detection on imbalanced data

Finds fraud in 284,807 transactions where only 0.17% are fraudulent. Four models were compared with and without SMOTE and tracked in MLflow. The deployed random forest reaches AUPRC 0.876 and ROC-AUC 0.970 behind a FastAPI service and a Streamlit dashboard.

  • Python
  • FastAPI
  • scikit-learn
  • XGBoost
  • MLflow
  • Streamlit
See more ↗ Live demo ↗
Crop disease detector predicting bacterial spot on a pepper leaf with a Grad-CAM heatmap

Crop Disease Detector

Leaf disease classification with explainable AI

Classifies 15 tomato, potato and pepper leaf conditions from a photo. Two-phase transfer learning raised ResNet18 test accuracy from 90.8% to 99.1%, and Grad-CAM heatmaps show what the model looked at. The images are lab-style, so accuracy on field photos will be lower.

  • Python
  • PyTorch
  • ResNet18
  • EfficientNet
  • MLflow
  • Gradio
See more ↗ Live demo ↗

About me

I’m a third-year B.Tech student in Data Science and Artificial Intelligence in New Delhi. I enjoy making AI systems work end to end, and I want to become the engineer who builds the backend that makes them reliable.

My READMEs say what each project does well and where it falls short.

More about me →
Backend implementation in VS Code
Real device testing
GitHub commits and contributions

Skills

AI and LLM

Retrieval, agents and speech

  • LangChain
  • LangGraph
  • Pinecone
  • Hugging Face
  • Cohere reranking
  • Whisper
  • RAGAS evaluation
  • Hybrid search

Machine learning

Models and explainability

  • PyTorch
  • scikit-learn
  • XGBoost
  • LightGBM
  • Transfer learning
  • SHAP
  • Grad-CAM
  • NumPy
  • Pandas

Backend and APIs

Services that serve the models

  • FastAPI
  • REST APIs
  • WebSockets
  • Pydantic
  • Streamlit
  • Gradio
  • Protocol design

Deployment and MLOps

Shipping and tracking

  • Docker
  • Render
  • Hugging Face Spaces
  • MLflow
  • Weights & Biases

Languages and tools

What I build with

  • Python
  • C++
  • Jupyter
  • Git
  • GitHub
  • Linux

Learning now

Backend foundations, in progress

  • PostgreSQL
  • SQLAlchemy
  • Redis
  • pytest
  • GitHub Actions
  • Data structures and algorithms

Additional projects

iTantra

A small prototype that sends speech as compact text over Wi-Fi, Bluetooth or LoRa radio. Built for Smart India Hackathon 2026.

  • Kotlin
  • Python
  • C++
GitHub ↗ Download APK ↗

Neural network from scratch

A neural network written only with NumPy, reaching 97.4% accuracy on MNIST.

  • Python
  • NumPy
GitHub ↗

DistilBERT sentiment fine-tune

DistilBERT fine-tuned on IMDB reviews, reaching 93.6% F1, with runs tracked in Weights & Biases.

  • Python
  • Hugging Face
  • Weights & Biases
GitHub ↗

Education

B.Tech in Data Science and Artificial Intelligence

Dr. Akhilesh Das Gupta Institute of Professional Studies, New Delhi (affiliated to GGSIPU)

August 2024 to July 2028

9.23 / 10 CGPA

Participation

Smart India Hackathon 2026

Cleared the internal round with iTantra, built for ISRO’s problem statement on Indian multilingual speech-to-text and text-to-speech radio access (PS 26173).

Let’s build something together.

I’m looking for an internship where I can work on a real backend.